⚡
Shallow-buffer first
KB-scale VOQs with deterministic drain. No multi-millisecond queueing tax, ever. Optimised for synchronous all-reduce, not bursty web traffic.
prototype Heads up: F-BOX NOS is an aspirational,
in-progress open-source project. Everything on this page is
for entertainment and discussion only — no shipping
binary, no production support, no vendor commitments.
v0.0.1-aspirational · pre-alpha · vibes only
One NOS
📰 We’re still aspirational, but somehow we’ve made press — thesis cited in SDxCentral.
One NOS
for the AI
fabric era.
F-BOX NOS is an open-source Network Operating System purpose-built for the kind of networks that train and serve frontier models — shallow-buffer, telemetry-driven, RDMA-aware, and unapologetically anti-bufferbloat.
Deep-buffer chassis switches were designed for the long-haul internet, not for collective communication across 100,000 GPUs. F-BOX takes the opposite bet: drain fast, signal early, route adaptively, and let the endpoints react in microseconds.
Figure 1 — A canonical AI training fabric, F-BOX top to bottom. No deep buffers; congestion is signalled, not absorbed.
The problem
Deep-buffer switches are a disaster for AI networking.
The packet-switched internet was tuned for human-scale traffic — bursty, lossy, latency-tolerant. Modern AI fabrics are the opposite: synchronous all-reduce collectives across tens of thousands of GPUs, where one straggler stalls the entire job.
Vendors who grew up selling long-haul gear have been pitching deep-buffer chassis switches as the answer. The argument is intuitive: more buffer means fewer drops. In an AI fabric the math inverts. A queue that takes milliseconds to drain is a queue that has already starved a synchronized collective.
Deep buffers: the trap
A full buffer must drain back out the same pipe that filled it, inflating effective RTT by 3× and stalling synchronous collectives. As Nvidia's networking lead put it, this was “a disaster after disaster after disaster” the last time it was tried at scale.
F-BOX: shallow + telemetric
F-BOX keeps buffers tiny on purpose. Congestion is detected immediately and signalled out-of-band (telemetry) and in-band (ECN/PFC/CNP), while flowlet-level adaptive routing re-balances the fabric in microseconds.
Architecture
A clean stack from silicon to scheduler.
F-BOX is structured as a thin, auditable control plane on top of a SAI-compliant hardware abstraction layer, with a dedicated AI-aware forwarding plane and a programmable telemetry fabric that any scheduler — Kubernetes, Slurm, Ray — can subscribe to.
What's in the box
State-of-the-art features, on day one (aspirationally).
The feature set below is what F-BOX targets at GA. Some pieces already have working prototypes; many are TODO. Status badges below each card.
📡
In-band telemetry (INT)
Every packet can carry per-hop latency, queue depth, and path fingerprint. Subscribe over gNMI/Kafka and feed it straight into your scheduler.
prototype🧭
Flowlet adaptive routing
ECMP without the elephant-flow tax. Per-flowlet hashing reacts to congestion in sub-millisecond windows, keeping every spine link busy.
in design🛡️
Lossless RDMA, sanely
RoCEv2 with auto-tuned PFC headroom, DCQCN, and a built-in PFC-storm detector. Watchdog reverts to lossy with a single intent flip.
in design🧱
SAI-native HAL
Runs on any Switch Abstraction Interface-compliant ASIC. Tomahawk, Spectrum, Silicon One, Teralynx — pick your silicon, keep your NOS.
prototype🦀
Memory-safe control plane
Written in Rust, with eBPF datapath hooks for in-kernel telemetry. No more CVE-of-the-week from a 1990s C codebase you inherited.
prototype🧩
Declarative intent API
Describe the fabric you want as YAML or gRPC. F-BOX reconciles continuously, like Kubernetes for the underlay.
in design🔭
Collective-aware QoS
NCCL and MPI ranks tag traffic; F-BOX pins all-reduce, all-gather, and reduce-scatter to dedicated traffic classes with bounded jitter.
TODO🌐
Scale-across, not just scale-out
Inter-DC fabrics with SRv6, EVPN, and BGP-LS. Train across two buildings as if they were one rack — without deep-buffer chassis gear in between.
TODOEcosystem
One NOS to subsume them all.
The open NOS world is brilliant, prolific, and hopelessly fragmented. SONiC, FBOSS, DENT, Stratum, Open Network Linux, Cumulus, OpenSwitch, FRR — every project nails a piece of the puzzle and reinvents the rest. F-BOX's aspirational goal: a single, AI-fabric-first NOS that absorbs the good ideas, ships compatibility shims, and lets the rest retire honourably.
Compatibility, not extinction. F-BOX ships shims for SONiC YANG models, FBOSS thrift interfaces, DENT switchdev assumptions, and Stratum P4Runtime endpoints. Migrate incrementally; don't fork your ops team.
Upstream what's good. F-BOX contributors commit to upstreaming AI-fabric improvements (telemetry, congestion control, RDMA hygiene) to FRR, SAI, and the Linux kernel so the whole ecosystem benefits — even the projects we ultimately subsume.
Roadmap
An honest, aspirational timeline.
This is the plan, not a promise. Dates are vibes. PRs welcome.
- SHIPPED
Q1 — Manifesto & reference topology
Public design doc, reference fabric topology, “why not deep buffers” whitepaper, and this website.
- IN PROGRESS
Q2 — SAI HAL + shallow-buffer datapath
Boot F-BOX on a Tomahawk 5 reference design; demonstrate measurable tail-latency wins versus a deep-buffer baseline on a 64-GPU all-reduce.
- IN PROGRESS
Q3 — INT + DCQCN + PFC watchdog
End-to-end in-band telemetry, auto-tuned DCQCN, and a PFC-storm detector with intent-driven recovery. SONiC YANG compatibility shim.
- PLANNED
Q4 — Flowlet adaptive routing
Sub-millisecond hash re-pinning per flowlet, with collective-aware QoS hints from NCCL and MPI ranks.
- PLANNED
Q5 — Scale-across
SRv6 + BGP-LS multi-DC fabric. Train across two buildings, no deep-buffer middleboxes required.
- DREAM
Year 2 — 1.0, CNCF sandbox proposal
Production-pinned 1.0, hardware bring-up guides for four ASIC families, and a path into the CNCF as the canonical open AI-fabric NOS.
“Buffer is not bandwidth. Buffer is borrowed time you have to pay back, with interest, on the return trip.”
— F-BOX design principle #1
Why this project exists
The AI-fabric debate has been shaped by two camps: vendors selling chassis switches with hundreds of megabytes of buffer per port, and vendors selling shallow-buffer Ethernet with telemetry-based congestion control. The first camp's pitch is intuitive. The second camp's pitch is correct. F-BOX exists to make the second camp's approach open, portable, and not owned by any single silicon vendor.
We don't expect to win every workload. We expect to win the ones that matter for the next decade of model training and high-throughput inference — and to make sure the answer to “which NOS should I run on my AI cluster?” isn't a vendor lock-in.
Join the (aspirational) community.
F-BOX NOS is a thought experiment that wants to grow into a real project. Reading code, writing design docs, arguing about buffer sizing on Mastodon — all of it counts. If you've shipped a NOS, debugged a PFC storm at 3am, or tuned DCQCN by hand, we want your opinions in the open.
Reminder: F-BOX NOS is an aspirational, in-progress project. This site is for entertainment and discussion only. Don't run it in production. Don't show it to your CTO. Or do — just bring snacks.